Demo: Verify, Don’t Trust: Open-Source AI Is Not Enough

Artificial intelligence (AI) is increasingly used in critical domains, yet models are often executed by third par-ties-making open-source access alone insufficient for establishing trust in results. We propose a design for an AI model registry that provides the foundation for verifiability of the execution by supporting the full model lifecycle, from model registration to inference validation. Our implementation leverages Ethereum, IPFS, and two verification approaches: zero-knowledge machine learning (ZKML), and optimistic ML (OPML) via dispute resolution. The system enables users to confirm both the identity of the model and the correctness of its execution, enhancing trust in open-source AI.